Online car-hailing multi-source data space-time situation visual analysis method

Abstract

The invention relates to an online car-hailing multi-source data space-time situation visual analysis method. The method comprises the following steps: selecting online car-hailing public original data in a certain region within a determined duration; dividing the area into an x * y grid matrix; preprocessing the trajectory data set in the original data and calculating the average vehicle speed Vn in each trajectory data; defining a time step length T, and calculating six traffic situation characteristics of each small area in the area under each time step length; constructing a regional feature matrix and a traffic map matrix; calculating a situation value feature embedding representation and a situation region association feature embedding representation, and respectively clustering; performing data processing on the clustering results to obtain respective corresponding grade matrixes; and carrying out dimension reduction processing on the grade matrix by adopting a t-sne method, and inputting the grade matrix into a visual interaction interface for display.

Description

Technical Field

[0001] The present invention relates to the field of point of interest situation analysis methods, and in particular to a visual analysis method of the spatiotemporal situation of multi-source data of online car-hailing. Background Art

[0002] In digital city management, online ride-hailing, as an important part of smart cities and sharing economy, has improved people's travel experience and provided data support for urban traffic management. However, online ride-hailing is currently facing the problem of imbalance between supply and demand. Analyzing people's travel activities from a macro perspective is crucial to alleviate the pressure in high-order areas. However, in addition to being affected by traffic participants, online ride-hailing activities are also affected by various external factors (such as POI). Situational awareness (SA) analysis is a comprehensive analysis process that aims to understand and predict the state, dynamics, and future trends in a complex environment or system. It is a method for analyzing multi-factor online ride-hailing situations. A complete SA analysis includes three aspects: perception, understanding, and prediction.

[0003] At present, more and more researchers have begun to pay attention to the situation analysis of online ride-hailing. In most studies, traffic elements are usually used as indicators to measure traffic situation, such as speed, density, flow, etc. However, situation is a relatively abstract concept that involves a comprehensive understanding and analysis of the environment, scenes, and events. Therefore, these single elements cannot fully and accurately reflect the traffic situation. In addition, most studies only focus on one aspect of situation analysis, especially on demand forecasting. In general, there is a lack of comprehensive quantitative indicators and correlation analysis to understand the situation of online ride-hailing. In addition, due to the high dimensionality and spatiotemporal characteristics of online ride-hailing data, it is difficult to understand the characteristics and changing trends of online ride-hailing data by only using a certain aspect of the situation method.

[0004] Visualization plays a vital role in the analysis of urban spatiotemporal data. It can reveal spatiotemporal patterns, analyze spatiotemporal trends, display spatial distribution, and analyze hot spots and events. Online ride-hailing data is a type of urban spatiotemporal data involving multiple factors. However, many studies only focus on a single data source, while ignoring the diversity and relevance of the data. In addition, compared with other transportation data, online ride-hailing data is highly dynamic and heterogeneous in the service field, and contains user order information, resulting in simple charts and static map visualizations that cannot effectively display the multidimensional characteristics and complex correlations of the data. Therefore, a special visualization method is needed to analyze the association between regions. Summary of the invention

[0005] In view of the above problems existing in the prior art, the technical problem to be solved by the present invention is: how to more accurately predict the changing trend of the online car-hailing situation.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solution: a method for visually analyzing the spatiotemporal situation of multi-source data of online car-hailing, comprising the following steps:

[0007] S1: Select the public raw data of online ride-hailing vehicles in a certain W area and W' time period, including order data set, trajectory data set and POI data set. Each order data in the order data set includes ID, timestamp, departure place and destination; each trajectory data in the trajectory data set includes ID, timestamp, longitude and latitude; each group of data in the POI data set includes category, longitude and latitude; the data duration can select data information in any time interval according to needs.

[0008] S2: Preprocess the trajectory data set: delete the trajectory data less than t minutes in the trajectory data set, generally t < 3 minutes; at the same time, use the window-based spline interpolation method to complete the deleted trajectory data, which is an existing technology; expand the trajectory data set and calculate the average vehicle speed V in each trajectory data n , where n represents the nth car;

[0009] S3: Divide the W area into x×y grids. Generally, each grid is defined as a square grid of 0.5km×0.5km, and each grid represents a small area.

[0010] S4: Define the time step T, calculate the six traffic situation characteristics of each small area corresponding to each grid in the W area at each time step. The time step defined here can be selected according to the data form, such as: the time step can be set to 15min, 30min, 60min, etc.; Among them, the six traffic situation characteristics include: regional average speed G(i,j) V 、Regional density G(i,j) K 、Regional flow G(i,j) Q , Regional situation index G(i,j) Is , Regional order departure quantity grid and regional order destination quantity grid The specific expressions are as follows:

[0011]

[0012] Where S represents the area of ​​the small area corresponding to the grid, N represents the number of vehicles passing through the unit time step T, i represents the row index in the W area and 0≤i≤x, j represents the column index in the W area 0≤j≤y, G(i,j) represents the small area corresponding to the (i,j)th grid, α is a small parameter to avoid the denominator being 0, which is set to 0.000001 here;

[0013] When the speed is 0, the situation index is set to 0, indicating no movement; conversely, the situation index can be used to judge the situation in the area: if the vehicle density is high and the traffic flow is large, a decrease in speed will increase the situation index, indicating that congestion may occur; conversely, a lower value indicates a relatively stable traffic condition.

[0014] Using the order data in a small area, the order volume in and out of the small area is counted and calculated. and The expression is as follows:

[0015]

[0016]

[0017] in, Represents the starting point of the e-th order; represents the destination of the e-th order;

[0018] S5: Constructing regional feature matrix: The six traffic situation features of all small areas are combined into a matrix, and then the maximum and minimum normalization method is used to normalize the matrix, thereby obtaining the regional feature matrix W;

[0019] S6: The regional feature matrix of W is embedded with the situation value feature through module P, and the module P is trained to obtain the trained module P'; GraphSAGE represents graph neural network, and MLP represents multi-layer perceptron, both of which are existing technologies.

[0020] S7: Construct a traffic map matrix: Each small area corresponds to an actual geographical area within the W area, and a small area is defined as a geographical node. Each geographical node has a unique ID tag. Orders whose departure and destination are both within the W area are defined as intra-regional orders, and the departure and destination of intra-regional orders are defined as order nodes. If two geographical nodes are physically adjacent, the two geographical nodes are considered to be geographically adjacent. If two order nodes correspond to the departure and destination of orders within the same region, the two order nodes are considered to be order-adjacent. All geographically adjacent geographical nodes are connected in pairs to form geographical edges, and order-adjacent order nodes are connected in pairs to obtain order edges. All geographical edges constitute the geographical edge set E1, and all order edges constitute the order edge set E2.

[0021] E1 and E2 form a traffic map matrix;

[0022] The maximum and minimum normalization method is used to normalize the features of each dimension of the traffic map matrix;

[0023] The two adjacent geographical nodes refer to two adjacent actual geographical areas. Specifically, the two adjacent regional nodes represent two geographical areas that are physically adjacent or closely connected. The geographical edge represents the relationship between these geographical areas. The two adjacent order nodes refer to the order departure point and order destination of the order. Each order corresponds to the departure point and destination of a specific order. Therefore, the adjacent order nodes refer to the connection between the departure point and destination of the same order. The order edge reflects the movement path of the order from one location to another in time and space.

[0024] The specific operation process of the connection is as follows: define the geographic edge set E1 and define the order edge set E2: traverse all nodes in area W. If the nodes G(i,j) and G(i',j') are geographically adjacent, add the edge [G(i,j), G(i',j')] to the set E1; if the nodes G(i,j) and G(i',j') are order adjacent, add the edge [G(i,j), G(i',j')] to the set E2.

[0025] E 1 ={[G(i,j),G(i′,j′)]|G(i,j) and G(i′,j′) are geographically adjacent}

[0026] Calculation formula: E 2 ={[G(i,j),G(i′,j′)]|G(i,j) and G(i′,j′) are adjacent in order} (11).

[0027] S8: The normalized traffic map matrix is ​​processed by module Q to obtain the situation area correlation feature embedding representation, and the module Q is trained to obtain the trained module Q';

[0028] S9: clustering calculations are performed on the situation value feature embedding representation and the situation region association feature embedding representation respectively using a Gaussian mixture model, where the Gaussian mixture model is an existing technology;

[0029] Among them, the situation value feature embedding means dividing the normalized traffic map matrix into several situation level matrices according to the similarity of the situation values; the situation area correlation feature embedding means dividing the normalized traffic map matrix into several area correlation level matrices according to the degree of correlation between the situation areas;

[0030] S10: Using the t-SNE method, the dimensionality reduction processing is performed on the several situation level matrices and the several regional correlation level matrices obtained in S9, and finally respective related two-dimensional matrices are obtained; the t-SNE method is a prior art method.

[0031] S11: For the area X to be tested, the regional feature matrix of X is constructed by the methods of S3-S5, and the regional feature matrix of X is input into the module P' trained by S6 to obtain the situation value feature embedding representation corresponding to X, and the traffic map matrix corresponding to X is constructed by the method of S6, and the traffic map matrix corresponding to X is input into the module Q' trained by S8 to obtain the situation region association feature embedding representation corresponding to X, and then the traffic map matrix corresponding to X is divided into several situation level matrices to be tested and several area association level matrices to be tested by the method of S9, and finally the method of S10 is used to obtain the two-dimensional matrices to be tested corresponding to several situation level matrices to be tested and several area association level matrices to be tested;

[0032] S12: Send the relevant two-dimensional matrix data to be tested obtained in S11 to the visualization interactive interface, so that the prediction result of the change trend of the number of online-hailing vehicles in area X can be displayed in the visualization interface.

[0033] Preferably, V in S2 n The calculation method is as follows:

[0034] The latitude and longitude of the departure and destination in each order are used as input, and the haversine formula is used to approximate the spherical distance between the two places. The formula is as follows:

[0035]

[0036] Where d is the spherical distance between the two locations, t n represents the travel time, r represents the radius of the sphere, represents the latitude of the departure and destination, λ 1 , 2 They represent the longitude of the departure and destination respectively, and both longitude and latitude are expressed in radians.

[0037] Preferably, the method for calculating the situation value feature embedding representation in S6 is as follows:

[0038] S6-1: Divide the regional feature matrix W obtained in S5 according to the time step T to obtain feature data of several time periods to form a feature data set;

[0039] S6-2: Construct module P. Module P includes GraphSAGE and MLP. GraphSAGE is composed of a two-layer convolutional neural network with LSTM as the aggregation function. GraphSAGE with LSTM as the aggregation function is the prior art.

[0040] S6-3: The feature data set obtained in S6-1 is used as the input of GraphSAGE, and the output is the situation prediction value matrix of region W. The obtained situation prediction value matrix is ​​subjected to nonlinear transformation and regression through MLP, and then the activation function is used to limit the value obtained after MLP processing to [0,1]. The activation function used here is Sigmoid

[0041] Previously, the formula for calculating the situation value was universal. The universal situation index only involved three basic traffic elements, which made it difficult to fully capture the situation characteristics of online ride-hailing vehicles. Therefore, this solution added order information features through modeling, so that the situation prediction value included 6 features of each node, which was richer than the universal situation index and achieved the final quantification of the situation of online ride-hailing vehicles.

[0042] S6-4: Set the learning rate of module P to 0.001, set the maximum number of training times, use the mean square error loss function MSELoss as the loss function to train module P, and update the parameters of model P according to the loss using the reverse update principle. When the training reaches the maximum number of training times, stop training and get the trained module P'. The calculation formula of MSELoss loss function is as follows:

[0043]

[0044] Among them, y i =G(i,j) Is , represents the situation index of the i-th small area, p i is the situation prediction value of the i-th small area;

[0045] S6-5: Use the regional feature matrix of W obtained in S5 as the input data of P' to obtain the embedded representation of the situation value feature of W.

[0046] Preferably, the method for calculating the embedded representation of situation region associated features in S8 is as follows:

[0047] S8-1: Construct module Q. Module Q includes GraphSAGE and negative sampling method. GraphSAGE is composed of a two-layer convolutional neural network with LSTM as the aggregation function. Negative sampling method is used to randomly construct negative sample edges to improve the learning and generalization ability of the model.

[0048] S8-2: The normalized traffic map matrix of the W region obtained in S7 is used as the input of module Q to obtain the situation region correlation feature embedding representation of the W region;

[0049] S8-3: Use the embedding method of mutual dot product to process the situation area correlation feature embedding representation of the W area to obtain the xy*xy regional correlation matrix; the area with higher correlation has a higher dot product score and is more likely to be associated, so as to simulate the correlation between urban area situations. If the correlation exceeds the set edge threshold, it is considered that there is an edge between the two traffic areas.

[0050] S8-4: Set the learning rate of module Q to 0.001, and set the maximum number of training times. Use the regional correlation matrix as the input of module Q, and use the binary cross entropy loss function BCEWithLogitsLoss as the loss function to train module Q. The binary cross entropy loss function combines the Sigmoid activation function with the binary cross entropy loss function (BCELoss). This can improve the computational efficiency of the model, and update the parameters of module Q based on the loss using the reverse update principle. When the training reaches the maximum number of training times, stop training and get the trained module Q'. The BCEWithLogitsLoss loss function calculation formula is as follows:

[0051]

[0052] Among them, u i represents the true label of the i-th small area, q i It indicates the edge prediction of the ith small area in module Q. Since it is a binary classification problem, its value is 1 if there is an edge between nodes, otherwise it is 0. If the prediction score is greater than the edge threshold, it is considered that there is an edge. The threshold is set in advance according to the actual situation. Otherwise, there is no edge, which is used as the basis for judging the existence of the edge.

[0053] The visual interactive interface includes a control panel, a traffic situation overview view, a traffic area information view, a regional POI view, a regional spatiotemporal stacking view, a traffic situation clustering view, and a traffic situation propagation view:

[0054] Control panel: Select a dataset and adjust parameters such as time period, number of samples, edge threshold, and time step.

[0055] Traffic situation overview view: provides a timeline to select a time point, showing the traffic flow view (the distribution of vehicles at that time point), situation heat map (the size of traffic element values ​​at that time point), order view (the link between the origin and destination of the vertex at that time point), POI distribution map (the location distribution of POIs on the map) and trajectory animation;

[0056] Traffic area information view: displays basic information of each traffic area, including area ID, number of orders in and out, speed / density / flow rate change trends and situation indicators;

[0057] Regional POI view: sort the number of POIs and count the distribution of POIs in the selected area;

[0058] Regional spatiotemporal stacking view: The regional spatiotemporal stacking view includes a spatial situation stacking view (displaying the cumulative situation value of each area throughout the day) and a time trajectory stacking view (displaying the trajectory distribution in a certain period of time in a certain area). Due to the lack of vehicle lane information, random trajectories have a lot of blank spaces and short and sparse trajectories. On the contrary, a large number of trajectories will lead to vertical expansion. Therefore, the present invention designs a trajectory layout method:

[0059] Trajectories are organized by their duration from short to long and arranged in layers in order: the first trajectory is placed in the first layer. Each subsequent trajectory is judged before being placed to see if there is a time intersection with the existing trajectory. If there is no intersection with any trajectory, it is placed in the first layer; if there is, it is assigned to the next layer of the intersection layer, and so on. At the same time, the trajectory block dynamically adjusts its height according to the traffic flow.

[0060] Traffic situation clustering view: Clusters areas according to attributes (situation feature embedding) and structure (correlation feature embedding), and displays the relevant attribute values ​​of the selected area in detail.

[0061] Traffic situation propagation view: Infer the situation propagation direction through regional attribute value differences and correlation edge links.

[0062] Compared with the prior art, the present invention has at least the following advantages:

[0063] 1. The fusion of multi-source data of online ride-hailing has been completed. Compared with the current reliance on a single data source, by integrating data from different sources, a more comprehensive traffic situation awareness can be obtained, thereby better capturing complex traffic dynamic changes.

[0064] 2. Compared with the holistic analysis of existing technologies, by calculating regional-level traffic attribute elements, it is possible to analyze the specific traffic conditions in different regions with higher resolution and pertinence. This regional analysis method helps optimize traffic scheduling and online car-hailing route planning. By integrating multi-source data and conducting regional-level analysis, it can provide more sophisticated decision support and help online car-hailing platforms better conduct operational scheduling, demand forecasting, and risk assessment.

[0065] 3. A general situation calculation method and a GraphSAGE-based method for quantifying the situation of online ride-hailing vehicles are proposed. This method is more reasonable than the traditional single indicator to measure the situation. Compared with a single indicator (such as a single traffic density or vehicle speed), this method can comprehensively measure multiple complex factors of traffic, such as traffic mobility, regional demand, vehicle status, etc., through situation calculation in multiple dimensions. Therefore, the situation analysis is more comprehensive and accurate.

[0066] 4. A situation value and regional correlation model was established based on GraphSAGE. Compared with the analysis based on a single graph theory model, GraphSAGE can effectively process graph data and is particularly suitable for dynamic and complex traffic network scenarios. GraphSAGE calculates the representation of each node by aggregating the information of neighboring nodes, which makes the model more robust and flexible when dealing with complex traffic networks and can better capture the laws of spatial and temporal changes.

[0067] 5. By establishing a situation value and regional correlation model, traffic areas can be clustered and classified to help online ride-hailing platforms customize services accurately in different traffic scenarios. For example, identify areas with high demand but insufficient services, deploy driver resources in advance, and improve platform efficiency. Through regional classification, differentiated solutions can be provided to help online ride-hailing platforms optimize resource allocation and scheduling.

[0068] 6. A stacking layout method for trajectory data lacking lane information is proposed to improve space utilization and prevent vertical space expansion. This stacking layout method can not only improve rendering efficiency, but also process trajectory data more flexibly in subsequent analysis, which can significantly improve the execution efficiency of the algorithm.

[0069] 7. It provides important methods and technical support for intelligent transportation systems. By integrating multi-source data and conducting visual analysis of spatiotemporal situations, it can not only improve the efficiency of traffic management, but also play an important role in decision-making, scheduling, and services. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The present invention discloses a visual analysis system for the spatiotemporal situation of multi-source data of online car-hailing;

[0071] Figure 2 The online car-hailing situation value prediction model disclosed in the present invention;

[0072] Figure 3 The regional correlation prediction model for online car-hailing disclosed in the present invention;

[0073] Figure 4 For the comparison between a single traffic element and situation indicators;

[0074] Figure 5 This is a comparative experiment between the GraphSAGE prediction model with LSTM as the aggregation function and the mainstream model;

[0075] Figure 6 This is the result of user research of the present invention. DETAILED DESCRIPTION

[0076] The present invention is described in further detail below.

[0077] The present invention proposes a visual analysis method of the spatiotemporal situation of multi-source data of online ride-hailing. By integrating online ride-hailing data (trajectory data, order data) from multiple sources, considering data diversity and correlation, a general traffic situation calculation method is proposed, and the online ride-hailing situation value is modeled by GraphSAGE to obtain the quantitative indicators of the online ride-hailing situation. At the same time, the regional correlation is modeled to analyze the correlation between regions. Finally, a complete visual analysis system of the spatiotemporal situation of multi-source data of online ride-hailing is formed.

[0078] See also Figure 1-Figure 6 , a method for visually analyzing the spatiotemporal situation of multi-source data of online car-hailing, comprising the following steps:

[0079] S1: Select the public raw data of online ride-hailing vehicles in a certain area W and a period W', including order data set, trajectory data set and POI data set. Each order data in the order data set includes ID, timestamp, departure place and destination; each trajectory data in the trajectory data set includes ID, timestamp, longitude and latitude; each group of data in the POI data set includes category, longitude and latitude;

[0080] S2: Preprocess the trajectory data set: delete the trajectory data less than t minutes in the trajectory data set; use the window-based spline interpolation method to complete the deleted trajectory data; expand the trajectory data set and calculate the average vehicle speed V in each trajectory data n , where n represents the nth car;

[0081] The S2 V n The calculation method is as follows:

[0082] The latitude and longitude of the departure and destination in each order are used as input, and the haversine formula is used to approximate the spherical distance between the two places. The formula is as follows:

[0083]

[0084] Where d is the spherical distance between the two locations, t n represents the travel time, r represents the radius of the sphere, represents the latitude of the departure and destination, λ 1 , 2 They represent the longitude of the departure and destination respectively, and both longitude and latitude are expressed in radians.

[0085] S3: Divide the W area into x×y grids, and each grid represents a small area;

[0086] S4: Define the time step T, and calculate the six traffic situation characteristics of each small area corresponding to each grid in the W area at each time step. The six traffic situation characteristics include: regional average speed G(i,j) V 、Regional density G(i,j) K 、Regional flow G(i,j) Q , Regional situation index G(i,j) Is , Regional order departure quantity grid and regional order destination quantity grid The specific expressions are as follows:

[0087]

[0088] Where S represents the area of ​​the small area corresponding to the grid, N represents the number of vehicles passing through the unit time step T, i represents the row index in the W area and 0≤i≤x, j represents the column index in the W area 0≤j≤y, G(i,j) represents the small area corresponding to the (i,j)th grid, α is a small parameter to avoid the denominator being 0, which is set to 0.000001 here;

[0089] Using the order data in a small area, the order volume in and out of the small area is counted and calculated. and The expression is as follows:

[0090]

[0091] in, Represents the starting point of the e-th order; represents the destination of the e-th order;

[0092] S5: Constructing regional feature matrix: The six traffic situation features of all small areas are combined into a matrix, and then the maximum and minimum normalization method is used to normalize the matrix, thereby obtaining the regional feature matrix W;

[0093] S6: The regional feature matrix of W is embedded in the state value feature through module P, and the module P is trained to obtain the trained module P';

[0094] The method for calculating the situation value feature embedding representation in S6 is as follows:

[0095] S6-1: Divide the regional feature matrix W obtained in S5 according to the time step T to obtain feature data of several time periods to form a feature data set;

[0096] S6-2: Construct module P. Module P includes GraphSAGE and MLP. GraphSAGE is a two-layer convolutional neural network with LSTM as the aggregation function.

[0097] S6-3: The feature data set obtained in S6-1 is used as the input of GraphSAGE, and the output is the situation prediction value matrix of the W region. The obtained situation prediction value matrix is ​​subjected to nonlinear transformation and regression through MLP, and then the activation function is used to limit the value obtained after MLP processing to between [0,1];

[0098] S6-4: Set the learning rate of module P to 0.001, set the maximum number of training times, use the mean square error loss function MSELoss as the loss function to train module P, and update the parameters of model P according to the loss using the reverse update principle. When the training reaches the maximum number of training times, stop training and get the trained module P'. The calculation formula of MSELoss loss function is as follows:

[0099]

[0100] Among them, y i =G(i,j) Is , represents the situation index of the i-th small area, p i is the situation prediction value of the i-th small area;

[0101] S6-5: Use the regional feature matrix of W obtained in S5 as the input data of P' to obtain the embedded representation of the situation value feature of W.

[0102] S7: Construct a traffic map matrix: Each small area corresponds to an actual geographical area within the W area, and a small area is defined as a geographical node. Each geographical node has a unique ID tag. Orders whose departure and destination are both within the W area are defined as intra-regional orders, and the departure and destination of intra-regional orders are defined as order nodes. If two geographical nodes are physically adjacent, the two geographical nodes are considered to be geographically adjacent. If two order nodes correspond to the departure and destination of orders within the same region, the two order nodes are considered to be order-adjacent. All geographically adjacent geographical nodes are connected in pairs to form geographical edges, and order-adjacent order nodes are connected in pairs to obtain order edges. All geographical edges constitute the geographical edge set E1, and all order edges constitute the order edge set E2.

[0103] E1 and E2 form a traffic map matrix;

[0104] The maximum and minimum normalization method is used to normalize the features of each dimension of the traffic map matrix;

[0105] S8: The normalized traffic map matrix is ​​processed by module Q to obtain the situation area correlation feature embedding representation, and the module Q is trained to obtain the trained module Q';

[0106] The method for calculating the embedded representation of the situation region associated features in S8 is as follows:

[0107] S8-1: Construct module Q, which includes GraphSAGE and negative sampling method. GraphSAGE is composed of two layers of convolutional neural network with LSTM as aggregation function;

[0108] S8-2: The normalized traffic map matrix of the W region obtained in S7 is used as the input of module Q to obtain the situation region correlation feature embedding representation of the W region;

[0109] S8-3: Process the embedded representation of the situational region correlation feature of region W by using the embedding mutual dot product method to obtain the regional correlation matrix xy*xy;

[0110] S8-4: Set the learning rate of module Q to 0.001, and set the maximum number of training times. Use the regional correlation matrix as the input of module Q, use the binary cross entropy loss function BCEWithLogitsLoss as the loss function to train module Q, and use the reverse update principle to update the parameters of module Q according to the loss. When the training reaches the maximum number of training times, stop training and get the trained module Q'. The BCEWithLogitsLoss loss function calculation formula is as follows:

[0111]

[0112] Among them, u i represents the true label of the i-th small area, q i Represents the edge prediction of the i-th small area in module Q.

[0113] S9: Use the Gaussian mixture model to perform clustering calculations on the situation value feature embedding representation and the situation region association feature embedding representation respectively;

[0114] Among them, the situation value feature embedding means dividing the normalized traffic map matrix into several situation level matrices according to the similarity of the situation values; the situation area correlation feature embedding means dividing the normalized traffic map matrix into several area correlation level matrices according to the degree of correlation between the situation areas;

[0115] S10: Using the t-SNE method, the dimensionality reduction processing is performed on the several situation level matrices and the several regional correlation level matrices obtained in S9, and finally the two-dimensional matrices related to each other are obtained;

[0116] S11: For the area X to be tested, the regional feature matrix of X is constructed by the methods of S3-S5, and the regional feature matrix of X is input into the module P' trained by S6 to obtain the situation value feature embedding representation corresponding to X, and the traffic map matrix corresponding to X is constructed by the method of S6, and the traffic map matrix corresponding to X is input into the module Q' trained by S8 to obtain the situation region association feature embedding representation corresponding to X, and then the traffic map matrix corresponding to X is divided into several situation level matrices to be tested and several area association level matrices to be tested by the method of S9, and finally the method of S10 is used to obtain the two-dimensional matrices to be tested corresponding to several situation level matrices to be tested and several area association level matrices to be tested;

[0117] S12: Send the relevant two-dimensional matrix data to be tested obtained in S11 to the visualization interactive interface, so that the prediction result of the change trend of the number of online-hailing vehicles in area X can be displayed in the visualization interface.

[0118] Experimental verification

[0119] Example: This experimental verification uses the online car-hailing order data (id, timestamp, departure place, destination) and trajectory data (id, timestamp, longitude, latitude) near the Chengdu Wuxiang Center on May 1, 2018 provided by an online car-hailing company, as well as the POI data (category, longitude, latitude) in the corresponding area obtained from Amap's open API.

[0120] The general situation calculation method proposed by the present invention is more reasonable than a single traffic factor, and it can reflect the areas with high and congested situations. For example, the speed in many areas is slow, but combined with the car density and traffic flow, it can be seen that the slow speed in some areas is not caused by congestion, but because there are fewer vehicles and slower speeds. Combining the three basic traffic indicators, the situation can extract the areas with high situations, which is verified in the visualization results, such as Figure 4 shown.

[0121] The present invention uses the graph embedding models GCN and GAT widely used in deep learning and GraphSAGE using different aggregation functions to conduct comparative experiments on the described real data sets. For the situation value prediction model (M1), the general evaluation method of the regression model is used, and the mean absolute error (MAE) is used as an indicator. For the regional situation correlation prediction model (M2), it is a binary classification task, and its performance is evaluated using accuracy.

[0122] To ensure the fairness of the experiment, the GCN and GAT models used in the present invention use the same number of layers and dimensionality definitions, that is, they all have 2 convolutional layers, the dimension of the regional situation association task is defined as (6×32×16), the dimension of the traffic situation prediction task is defined as (6×16×1), the data set is divided into training set, validation set and test set according to (8:1:1), GAT uses single-head attention, the model learning rate is set to 0.001, the parameters are adjusted using the Adam optimizer, the number of negative sampling times in situation association prediction is 3, the batch size in traffic situation prediction is set to 4, and a total of 50 rounds of training are performed until the model converges. The models constructed by the present invention all use GraphSAGE graph convolutional networks with LSTM as the aggregation function, and the results are as follows: Figure 5 shown.

[0123] The results show that the model of the present invention performs well in both situation value prediction and regional correlation prediction. This advantage may be attributed to the temporal nature of online car-hailing data, and the long short-term memory (LSTM) network has been shown to be better at processing temporal data. In addition, the GCN and GAT models are more suitable for training static graphs, while the GraphSAGE sampling idea has better dynamic applicability. The GraphSAGE model using GCN as an aggregator does not perform as well as GCN in situation correlation prediction because convolutional aggregation introduces the "over-smoothing" problem, resulting in information loss.

[0124] To further verify the effectiveness of the method, the present invention uses a Likert scale to construct a questionnaire to evaluate user experience and collect feedback. The survey had 13 participants, including 5 researchers in the field of visualization and visual analysis, 2 researchers in the field of software engineering, 2 researchers in the field of machine learning, and 4 researchers in the field of transportation. According to the visualization analysis task, the following five tasks and questions were formulated:

[0125] Q1 analyzes the overall evolution of online ride-hailing traffic throughout the day;

[0126] Q2 investigates factors that contribute to high-situation conditions in specific areas;

[0127] Q3 evaluates the system's ability to identify traffic situation distribution patterns in the area;

[0128] Q4Do you think the visual design of the system conforms to general aesthetics?

[0129] Q5 Do you think the system's interactive design is reasonable and easy to operate?

[0130] Before the experiment, participants were given 15 minutes to learn about the dataset and how to use the system. Then, 30 minutes were allocated for free exploration, during which participants interacted with the system and recorded their feedback. After the experiment, participants completed a questionnaire (Q1 to Q5) where the Likert rating for each task ranged from 1 (not at all accepted) to 5 (completely accepted). Figure 6 As shown in the results, the system performs well in performing visual analysis tasks, with an average acceptance rate of about 80% and Likert scale scores of [4.15, 4.23, 4.00, 4.30, 3.92]. Among them, they believe that the situation quantification method is more reasonable than a single element to evaluate the regional traffic situation. At the same time, combined with machine learning methods to predict richer situations, integrating geographic spatial features and time features, it provides a new and effective method for shared travel traffic situation analysis.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A visual analysis method of spatiotemporal situation of multi-source data of online car-hailing, characterized by: The steps include: S1: Select the public raw data of online ride-hailing vehicles in a certain area W and a period W', including order data set, trajectory data set and POI data set. Each order data in the order data set includes ID, timestamp, departure place and destination; each trajectory data in the trajectory data set includes ID, timestamp, longitude and latitude; Each set of data in the POI dataset includes category, longitude and latitude; S2: Preprocess the trajectory data set: delete the trajectory data less than t minutes in the trajectory data set; and use the window-based spline interpolation method to complete the deleted trajectory data; Expand the trajectory data set and calculate the average vehicle speed V in each trajectory data n , where n represents the nth car; S3: Divide the W area into x×y grids, and each grid represents a small area; S4: Define the time step T, and calculate the six traffic situation characteristics of each small area corresponding to each grid in the W area at each time step. The six traffic situation characteristics include: regional average speed G(i,j) V 、Regional density G(i,j) K 、Regional flow G(i,j) Q , Regional situation index G(i,j) Is , Regional order departure quantity grid and regional order destination quantity grid The specific expressions are as follows: Where S represents the area of ​​the small area corresponding to the grid, N represents the number of vehicles passing through the unit time step T, i represents the row index in the W area and 0≤i≤x, j represents the column index in the W area 0≤j≤y, G(i,j) represents the small area corresponding to the (i,j)th grid, α is a small parameter to avoid the denominator being 0, which is set to 0.000001 here; Using the order data in a small area, the order volume in and out of the small area is counted and calculated. and The expression is as follows: in, Represents the starting point of the e-th order; represents the destination of the e-th order; S5: Constructing regional feature matrix: The six traffic situation features of all small areas are combined into a matrix, and then the maximum and minimum normalization method is used to normalize the matrix, thereby obtaining the regional feature matrix W; S6: The regional feature matrix of W is embedded in the state value feature through module P, and the module P is trained to obtain the trained module P'; S7: Construct a traffic map matrix: Each small area corresponds to an actual geographical area within the W area, and a small area is defined as a geographical node. Each geographical node has a unique ID tag. Orders whose departure and destination are both within the W area are defined as intra-regional orders, and the departure and destination of intra-regional orders are defined as order nodes. If two geographical nodes are physically adjacent, the two geographical nodes are considered to be geographically adjacent. If two order nodes correspond to the departure and destination of orders within the same region, the two order nodes are considered to be order-adjacent. All geographically adjacent geographical nodes are connected in pairs to form geographical edges, and order-adjacent order nodes are connected in pairs to obtain order edges. All geographical edges constitute the geographical edge set E1, and all order edges constitute the order edge set E2. E1 and E2 form a traffic map matrix; The maximum and minimum normalization method is used to normalize the features of each dimension of the traffic map matrix; S8: The normalized traffic map matrix is ​​processed by module Q to obtain the situation area correlation feature embedding representation, and the module Q is trained to obtain the trained module Q'; S9: Use the Gaussian mixture model to perform clustering calculations on the situation value feature embedding representation and the situation region association feature embedding representation respectively; Among them, the situation value feature embedding means dividing the normalized traffic map matrix into several situation level matrices according to the similarity of the situation values; the situation area correlation feature embedding means dividing the normalized traffic map matrix into several area correlation level matrices according to the degree of correlation between the situation areas; S10: Using the t-SNE method, the dimensionality reduction processing is performed on several situation level matrices and several regional correlation level matrices obtained in S9, and finally the two-dimensional matrices related to each other are obtained; S11: For the area X to be tested, the regional feature matrix of X is constructed by the methods of S3-S5, and the regional feature matrix of X is input into the module P' trained by S6 to obtain the situation value feature embedding representation corresponding to X, and the traffic map matrix corresponding to X is constructed by the method of S6, and the traffic map matrix corresponding to X is input into the module Q' trained by S8 to obtain the situation region association feature embedding representation corresponding to X, and then the traffic map matrix corresponding to X is divided into several situation level matrices to be tested and several area association level matrices to be tested by the method of S9, and finally the method of S10 is used to obtain the two-dimensional matrices to be tested corresponding to several situation level matrices to be tested and several area association level matrices to be tested; S12: Send the relevant two-dimensional matrix data to be tested obtained in S11 to the visualization interactive interface, so that the prediction result of the change trend of the number of online-hailing vehicles in area X can be displayed in the visualization interface.

2. The method for visually analyzing the spatiotemporal situation of multi-source data of online car-hailing as claimed in claim 1, characterized in that: The S2 V n The calculation method is as follows: The latitude and longitude of the departure and destination in each order are used as input, and the semi-segment formula is used to approximate the spherical distance between the two places. The formula is as follows: Where d is the spherical distance between the two locations, t n represents the travel time, r represents the radius of the sphere, It represents the latitude of the departure place and the destination, λ1 and λ2 represent the longitude of the departure place and the destination respectively, and both longitude and latitude are expressed in radians.

3. A method for visually analyzing the spatiotemporal situation of multi-source data of online car-hailing as claimed in claim 2, characterized in that: The method for calculating the situation value feature embedding representation in S6 is as follows: S6-1: Divide the regional feature matrix W obtained in S5 according to the time step T to obtain feature data of several time periods to form a feature data set; S6-2: Construct module P. Module P includes GraphSAGE and MLP. GraphSAGE is a two-layer convolutional neural network with LSTM as the aggregation function. S6-3: The feature data set obtained in S6-1 is used as the input of GraphSAGE, and the output is the situation prediction value matrix of the W region. The obtained situation prediction value matrix is ​​subjected to nonlinear transformation and regression through MLP, and then the activation function is used to limit the value obtained after MLP processing to between [0,1]; S6-4: Set the learning rate of module P and the maximum number of training times. Use the mean square error loss function MSELoss as the loss function to train module P, and update the parameters of model P according to the loss using the reverse update principle. When the training reaches the maximum number of training times, stop training and get the trained module P'. The calculation formula of MSELoss loss function is as follows: Among them, y i =G(i,j) Is , represents the situation index of the i-th small area, p i is the situation prediction value of the i-th small area; S6-5: Use the regional feature matrix of W obtained in S5 as the input data of P' to obtain the embedded representation of the situation value feature of W.

4. A method for visually analyzing the spatiotemporal situation of multi-source data of online car-hailing as claimed in claim 3, characterized in that: The method for calculating the embedded representation of the situation region association feature in S8 is as follows: S8-1: Construct module Q, which includes GraphSAGE and negative sampling method. GraphSAGE is composed of two layers of convolutional neural network with LSTM as aggregation function; S8-2: The normalized traffic map matrix of the W region obtained in S7 is used as the input of module Q to obtain the situation region correlation feature embedding representation of the W region; S8-3: Process the embedded representation of the situational region correlation feature of region W by using the embedding mutual dot product method to obtain the regional correlation matrix xy*xy; S8-4: Set the learning rate of module Q and the maximum number of training times. Use the regional correlation matrix as the input of module Q. Use the binary cross entropy loss function BCEWithLogitsLoss as the loss function to train module Q. Update the parameters of module Q using the reverse update principle based on the loss. When the training reaches the maximum number of training times, stop training and get the trained module Q'. The BCEWithLogitsLoss loss function calculation formula is as follows: Among them, u i represents the true label of the i-th small area, q i Represents the edge prediction of the i-th small area in module Q.

Images